Fig. 9.

a The likelihood (frequency) of convergence plot against the magnitude of a random perturbation to the ground-truth fitting results computed by tracking through a trinocular sequence. The results show that the calibrated multi-view algorithms are more robust than the uncalibrated multi-view algorithm discussed in Sect. 3, which itself is more robust than the 2D+3D single-view algorithm (Xiao et al. 2004a). b The rate of convergence is estimated by plotting the average error after each iteration against the iteration number. The results show that the calibrated multi-view algorithms converge faster than the uncalibrated algorithm, which converges faster than the single-view 2D+3D algorithm